An Oracle White Paper September 2010 Handling Memory Ordering in Multithreaded Applications with Oracle Solaris Studio 12 Update 2: Part 1, Compiler
Introduction... 1 What Is Memory Ordering?... 2 Compiler Memory Ordering... 2 Volatile Variables Are not Necessarily the Best Solution... 4 Complier... 5 Conclusion... 7
Introduction Oracle Solaris Studio 12 Update 2 introduces a new header file <mbarrier.h>. This header file provides a set of memory ordering intrinsics that are useful for enforcing memory ordering constraints in multithreaded applications. This article is part 1 of a two-part series. This part discusses how compiler barriers can be used to stop the compiler from generating code that is incorrect due to reordered memory accesses. Part 2 discusses how memory barriers or memory fences can be used to ensure that the processor does not reorder memory operations. 1
What Is Memory Ordering? Memory ordering is the order in which memory operations (loads and stores) are performed. There are two ways that memory ordering might be changed: The compiler might change memory ordering as a result of some compile-time optimizations. The processor might change memory ordering at run time. Compiler Memory Ordering For performance reasons, a compiler performs various optimizations. It might hold variable values in registers (rather than in memory), reorder expressions, avoid reloading values, eliminate the use of variables altogether, and so on. There are two constraints to what the compiler can do. First of all, it must produce code that is functionally equivalent to the code as written. This is the as-if rule, which basically says that the output of the optimized program should be as if the program were executed exactly as written. The second constraint is the user-specified flags passed to the compiler. If these flags say, for example, that pointers never alias, then the compiler can use that assumption in producing more optimal code. The code shown in Listing 1 reflects a situation in which a compiler could produce a more-optimal code sequence by reordering the memory operations. The code starts a team of threads. Each thread waits until it is signaled before starting work and then signals as soon as it completes work. The main thread is responsible for creating the team of threads, starting them, and then waiting until their work is complete. Listing 1. Multithreaded Code in which the Compiler Might Perform Undesirable Optimizations #include <stdio.h> #include <pthread.h> volatile int volatile int pthread_t start[10]; ended[10]; threads[10]; void * work( void * param) int id = (int) param; while( start[ id ] == 0 ) // Wait until work started printf( "Thread %i started\n", id ); ended[ id ] = 1; // Indicate that work completed printf( "Thread %i finished\n", id ); 2
void dowork() for (int count = 0; count < 10; count++ ) start[ count ] = 1; // Start thread working for (int count = 0; count < 10; count++ ) while( ended[ count ] == 0 ) // Wait until thread completes work int main() // Create team of threads for( int count = 0; count < 10; count++ ) start[ count ] = 0; ended[ count ] = 0; pthread_create( &threads[count], 0, work, (void*)count ); dowork(); // Join team of threads for( int count = 0; count < 10; count++ ) pthread_join( threads[count], 0 ); Compiling and running the code in Listing 1 without optimization produces the results shown in Listing 2. The threads start and finish in an unspecified order. Listing 2. Compiling and Running Code Without Optimization % cc -mt listing1.c %./a.out Thread 1 started Thread 1 finished Thread 9 started Thread 9 finished Thread 5 started Thread 5 finished Thread 0 started Thread 0 finished Thread 3 started Thread 3 finished Thread 4 started Thread 4 finished Thread 6 started Thread 6 finished Thread 7 started Thread 7 finished Thread 2 started Thread 2 finished Thread 8 started Thread 8 finished 3
We can compare the results in Listing 2 with the results from when the code is compiled with optimization (shown in Listing 3). Listing 3. Compiling and Running Code with Optimizations % cc -g -O -mt listing1.c %./a.out Thread 0 started Thread 0 finished Thread 1 started Thread 1 finished Thread 2 started Thread 2 finished Thread 3 started Thread 3 finished Thread 4 started Thread 4 finished Thread 5 started Thread 5 finished Thread 6 started Thread 6 finished Thread 7 started Thread 7 finished Thread 8 started Thread 8 finished Thread 9 started Thread 9 finished When the code is compiled with optimization, the threads start and complete in order. The reason for this is an optimization that the compiler performed on the routine dowork(). The optimized version of this routine is shown in Listing 4. The compiler combined the two loops in the dowork() routine. This resulted in the main thread starting each child thread and then waiting until that child thread completes. This is in contrast to the desired behavior of starting all the child threads in parallel. Listing 4. Compiler Optimized Version of the Routine dowork() void dowork() for (int count = 0; count < 10; count++ ) start[ count ] = 1; // Start thread working while( ended[ count ] == 0 ) // Wait until thread completes work Volatile Variables Are not Necessarily the Best Solution It is worth discussing the use of the keyword volatile in this code. The keyword volatile means that a value needs to be loaded from memory before use and stored back to memory immediately after use. This behavior is useful because it ensures that loops, such as the one in Listing 5, are not converted to infinite loops during optimization. 4
Listing 5. Loop that Could Be Converted to an Infinite Loop During Optimization while( ended[ count ] == 0 ) // Wait until thread completes work However, using volatile variables is not without costs. If the variable is used as part of a computation, its value cannot be carried in a register; it needs to be refreshed at each use. So there is some performance impact. The use of volatile variables does not make code correct. It essentially reduces the optimization of code that uses the variables. For example, the code in Listing 6 contains a data-race condition regardless of whether the variable is declared as volatile. Listing 6. Volatile Variables Do not Protect Against Data Races void increment(int * variable) (*variable)++; Finally, as we saw in the example from Listing 1, volatile variables do not protect against side effects of optimization. The conclusion we should draw from this is that volatile variables are not a panacea; they might solve a subset of problems, but they do not solve all of them. At this point it is appropriate to ask What is required in order for the code to function correctly? The answer to that is we need points in the code that the compiler cannot optimize around. If we take the code in Listing 5, we want the compiler to have to reload the variable ended[] at every iteration of the loop, and we do not want the compiler to assume that the variable is invariant. Similarly, we want the compiler to avoid merging the two loops, as shown in the optimization in Listing 4. So what we need is a compiler barrier. Complier A compiler barrier is a sequence point. At such a point, we want all previous operations to have stored their results to memory, and we want all future operations to not have been started yet. The most common sequence point is a function call. We can recode Listing 5 to use a function call to achieve the desired effect. This method is shown in Listing 7. The call to DoNothing() can immediately return having, as its name suggests, done nothing. However, the compiler does not know that nothing was done; the compiler has to assume that the function call might have changed global data. Consequently, the compiler needs to reload the value of the variable ended[]. Listing 7. Using a Function Call as a Compiler Barrier while( ended[ count ] == 0 ) DoNothing() // Wait until thread completes work 5
There are two problems with this approach. The first problem is that it uses an unnecessary function call, and the code might be more efficient if the function call were not performed. The second, more serious, problem is that under optimization, the call to DoNothing() might be in-lined. If this happens, the sequencing point is removed, and the loop is converted once again to an infinite loop. The header file <mbarrier.h> introduces the compiler_barrier() intrinsic, which is designed to provide the required functionality. This intrinsic has two advantages. The first is that the barrier remains during optimization, and the second is that it does not generate any instructions. The compiler_barrier() intrinsic can be used to provide the desired reloading of the variable ended[], as shown in Listing 8. Listing 8. Using the compiler_barrier() Intrinsic to Avoid Generating Infinite Loops During Optimization // Wait until thread completes work while( ended[ count ] == 0 ) compiler_barrier() The compiler_barrier()intrinsic can be used to avoid the requirement of declaring the variables volatile, and it can be placed between the two loops in the dowork() routine to ensure that the compiler does not merge them. The code modified to use the compiler barrier is shown in Listing 9. The changes are indicated in bold. Listing 9. Full Listing Modified to use the compiler_barrier() Intrinsic #include <stdio.h> #include <pthread.h> #include <mbarrier.h> int start[10]; // volatile removed int ended[10]; // volatile removed pthread_t threads[10]; void * work( void * param) int id = (int) param; // Use a compiler barrier to reload start[id] every iteration while( start[ id ] == 0 ) compiler_barrier(); // Wait until work started printf( "Thread %i started\n", id ); ended[ id ] = 1; // Indicate that work completed printf( "Thread %i finished\n", id ); 6
void dowork() for (int count = 0; count < 10; count++ ) start[ count ] = 1; // Start thread working // Use a compiler barrier to stop the two loops being merged compiler_barrier(); for (int count = 0; count < 10; count++ ) // Wait until thread completes work // Use a compiler barrier to reload ended[count] every iteration while( ended[ count ] == 0 ) compiler_barrier(); int main() // Create team of threads for( int count = 0; count < 10; count++ ) start[ count ] = 0; ended[ count ] = 0; pthread_create( &threads[count], 0, work, (void*)count ); dowork(); // Join team of threads for( int count = 0; count < 10; count++ ) pthread_join( threads[count], 0 ); Conclusion This article, which is part 1 of a two-part series, described how to use compiler barriers to stop the compiler from generating code that is incorrect due to reordered memory accesses. Part 2 of the series discusses how memory barriers or memory fences can be used to ensure that the processor does not reorder memory operations. 7
Handling Memory Ordering in Multithreaded Applications with Oracle Solaris Studio 12 Update 2: Part 1, Compiler September 2010 Author: Darryl Gove Oracle Corporation World Headquarters 500 Oracle Parkway Redwood Shores, CA 94065 U.S.A. Worldwide Inquiries: Phone: +1.650.506.7000 Fax: +1.650.506.7200 oracle.com Copyright 2010, Oracle and/or its affiliates. All rights reserved. This document is provided for information purposes only and the contents hereof are subject to change without notice. This document is not warranted to be error-free, nor subject to any other warranties or conditions, whether expressed orally or implied in law, including implied warranties and conditions of merchantability or fitness for a particular purpose. We specifically disclaim any liability with respect to this document and no contractual obligations are formed either directly or indirectly by this document. This document may not be reproduced or transmitted in any form or by any means, electronic or mechanical, for any purpose, without our prior written permission. Oracle and Java are registered trademarks of Oracle and/or its affiliates. Other names may be trademarks of their respective owners. AMD, Opteron, the AMD logo, and the AMD Opteron logo are trademarks or registered trademarks of Advanced Micro Devices. Intel and Intel Xeon are trademarks or registered trademarks of Intel Corporation. All SPARC trademarks are used under license and are trademarks or registered trademarks of SPARC International, Inc. UNIX is a registered trademark licensed through X/Open Company, Ltd. 0310